The MinGW numbers seem lower than I would expect, given the No VPN stats.
It might be instructive to try the test with --cipher none --auth none to
factor out crypto overhead.
I just ran the same tests with these settings and the numbers only varied
by about +/- 20K.
I'm also surprised that gcc and the MS compilers would show such large
differences. I would have thought that the bottlnecks would be low level
crypto and the context switching to and from the TAP-Win32 driver.
I know that a lot of work went into mingw to make the gcc environment work
on the win32 platform. A quick google for compiler benchmarks shows that
overall, MSVC and ICL outperform MINGW on the windows platform. Perhaps it
is nothing more than a compiler flag missing?!? I do know that for a while
I was running the MINGW version for several windows to windows tunnels and
had noticeable cpu usage. When I switched all the tunnels to the MSVC or
ICL builds, the cpu usage dropped greatly.
I have posted my builds of Beta20 to http://www.digital-z.net/openvpn and
encourage others to download them and run a few tests. Maybe it is just my
Openvpn-users mailing list